52 interviews and a 673-user survey yield HAABI, a 20-item scale for human-AI affective bonding

Understanding the Rising Human-AI Affective Bonding: Conceptualization and HAABI Scale Development

Lu Chen, Xiaoran Xue, Rongqi Ding, Fenghua Tang, Anji Zhou, Chenxi Wang, Mengyu Miranda Gao, Zhuo Rachel Han

cs.HC

2026-05-28

BNU built HAABI from 52 interviews and a 673-person survey: a 20-item scale splitting human-AI bonding into emotional realism, separation anxiety, investment, and romantic intimacy.

What problem this solves

Companion AI needs no scale argument: by July 2025 the apps had accumulated 220 million downloads on the Apple App Store and Google Play, with 60 million in the first half of 2025 alone, up 88% year over year. Both sides of the debate over whether AI companionship relieves loneliness or manufactures dependence have studies behind them, but the instruments underneath mostly adapt interpersonal attachment scales, swapping "partner" for "AI" and keeping the anxiety-avoidance structure. They measure how much an AI bond resembles a human one, not what the bond itself consists of.

A psychology team at Beijing Normal University treated that measurement gap as the actual research question. The goal was not to rule on whether these bonds are good or bad, but to describe the construct from users' own accounts and turn it into a validated scale.

Method

Two sequential studies, following standard scale-development procedure:

The six qualitative dimensions collapsed into four factors, and 49 items shrank to 20. The collapse is informative in itself: perceived agency and perceived relationship merged into emotional realism, because "the AI has a mind of its own" and "this relationship is real" are hard to separate in lived experience. Emotional projection split into separation anxiety on one end and emotional investment on the other.

Results

MetricResult
Factor structure20 items, 4 factors, 47.64% cumulative variance
CFA fit (independent sample)CFI = .919, TLI = .906, RMSEA = .056, SRMR = .062
ReliabilityTotal α = .905, ω = .927; subscale α = .763-.817
Known-groups validityInstrumental 2.23 < confidant 3.59 < relational 4.17, η² = .62
Criterion correlationsAttachment anxiety r = .54, AI-related flourishing r = .53, emotional-interaction proportion r = .44
Incremental validityΔR² = .102 (β = .43) for AI-related flourishing beyond attachment and use measures

The scale tracks interaction quality rather than volume: it correlates with the proportion of emotional interaction (r = .44) far more than with daily interaction time (r = .27). Scores rise monotonically across the three user groups, with all pairwise differences significant.

One demographic pattern stands out: users with more satisfying prior romantic relationships scored higher on emotional realism (r = .14) and lower on separation anxiety (r = −.13). Offline relationship status showed no association with the total score; the effects live in specific subscales. Roleplay and companion-app users outscored general-assistant users (d = 0.48), and men scored higher than women on emotional realism (d = .33) and romantic intimacy (d = .21).

Why it matters

For companion-product teams, this is a validated Chinese-language instrument that decomposes bonding into four subscales, so segmentation can separate intensity from composition: emotional realism for work on perceived authenticity, romantic intimacy for AI-romance features.

The separation-anxiety dimension ties platform decisions to psychological consequences. The paper cites the genuine grief in user communities when OpenAI retired GPT-4.0 in 2025: model updates, memory wipes, and shutdowns are unilateral terminations of a relationship, not feature changes. There is now an individual-level measure for pricing that risk.

To be clear about what this is: a psychometrics paper. No model is improved; the contribution is the instrument.

Limitations

The authors list three: the sample is entirely Chinese-speaking adults, and cultural differences in emotional norms may not preserve the factor structure; the cross-sectional design cannot establish causal or developmental order; everything is self-report, with social-desirability risk highest on the romantic and sexual items.

A few more from reading closely. Study 1 drew its framework from a sample that was 86.5% female with a mean age of 23. The known-groups comparison sorts users by self-reported relationship type and then tests whether scores differ along that same self-report, which builds circularity into the validity evidence. Incremental validity for perceived AI emotional capability was ΔR² = .009, barely above noise. Four factors explain 47.64% of variance, leaving more unexplained than explained, and the narrative connecting six qualitative dimensions to four factors reads as post hoc rationalization more than derivation.

Terms

Source

What people are saying

Related papers

All paper explainers